Training and the knowledge cutoff
A chatbot learned from text up to a certain date. After that date, it knows nothing. Like a newspaper from last year. What's in it was right then, but yesterday's news isn't in it. Some chatbots can search the internet to fill the gap.
First, what do you think?
You ask a chatbot who won last night's match. Can it know? Have a guess.
What is the knowledge cutoff?
The whole story in three sentences. A chatbot learned from a huge pile of text. That pile stops at a certain date. What happened after that date, the predictor never saw.
Think of a newspaper from last year. Everything in it was right when it was printed. But yesterday's news isn't in it. You can read it as often as you like. Yesterday's news won't appear.
The knowledge cutoff is that date: the moment the training text stops.
- Training comes first. The model learns from text before you ever use it. That takes an enormous amount of text.
- Then the pile is closed. One maker lists a training data cutoff for each model. That's the date its training text stops.
- The last months are thin. The same maker also lists a reliable knowledge cutoff. That's the date up to which its knowledge is most complete.
- Your chat doesn't teach it. What you type goes on the whiteboard, not into the training.
Some chatbots can search the internet. One maker says search lets a model answer with up-to-date information beyond its cutoff. The search results go on the whiteboard. Then the predictor can use them.
A first small step: is it about something recent? Ask the chatbot to search and show its sources.
What the predictor does
The predictor only knows the patterns in its training text. That's the rule.
Ask about last night's match without search. The predictor has no result in its pile. It may still predict an answer that sounds right. A name, a score. That answer isn't news. It's a guess dressed up as news.
With search, it works differently. The app looks up pages and puts them on the whiteboard. The predictor reads them like a pasted document. One maker says the answer then comes with citations: links to where it found things.
When does a chatbot search? One maker gives a list. It searches for recent news, current prices, scores, and things that change. For stable facts, maths or writing a poem, it answers without searching.
The predictor knows its pile, up to the cutoff. For anything newer, it needs to search, or it guesses.
Try it yourself
Old newspaper or fresh search? Sort the questions. Which can the predictor answer from its training, and which need a search?
Everything runs in your browser. Nothing you type is sent anywhere.
Use the free version of any chatbot. Don't type anything personal.
Ask a chatbot:
What is your knowledge cutoff? And can you search the internet in this chat?What to expect. Many chatbots name a month or year. Some say they can search.
What to look for. Now ask about something from this week. Does it show sources? Click one, and check that it really says that.
In practice
Three people read last year's newspaper without knowing it.
Case 1The opening hours
Daan (14) asks a chatbot when the swimming pool opens on Sunday. It gives a time. When he arrives, the pool is closed.
Decide that chatbots are useless.
Opening hours change. Without search, the predictor only had old text, or guessed.
Check the pool's own website. Or ask the chatbot to search, and click the source.
The source is today's newspaper. The training is last year's.
Case 2The new rule
Priya (38) asks a chatbot about a tax rule for her small business. The answer is clear and confident. The rule changed this year.
Trust it, because it sounds sure.
Sounding sure tells you nothing about the date. The predictor writes the same way about old and new.
Ask which year the answer is about. Check it on the tax authority's website.
Rules change. The official site has the current version.
Case 3The history homework
Joost (12) wants to know how the pyramids were built. He's not sure the chatbot can help.
Think chatbots only know old stuff, so it's all outdated.
For stable knowledge, the cutoff hardly matters. The pyramids didn't change last month.
Use the chatbot to explain. Check key facts in a book or on a trusted website.
The cutoff matters for news and things that change. Not for everything.
By situation
For history, maths or how things work, the cutoff hardly matters. For current events, use a chatbot that searches, and check its sources.
Prices, rules and product details change. Ask the chatbot to search, or check the official source yourself.
Opening hours, timetables, the weather, the score: these change every day. Look them up on the source itself, or let the chatbot search.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| A chatbot knows everything on the internet | No | It knows its training text, up to a date. Some can search, but that's a separate step. |
| It learns from my chat, so tomorrow it knows | No | Your chat goes on the whiteboard, not into its training. |
| If it sounds sure, it's up to date | No | It sounds the same about old and new things. |
| The cutoff makes everything outdated | No | Stable knowledge, like maths or history, hardly changes. |
Related topics
Read first
- The context windowthe whiteboard where search results land.
Next
- Hallucinationswhy it sometimes makes things up with confidence.
Related
- Search and researchAI that searches the web for you.
Sources
- documentationAnthropic. Models overview.Each model has a training data cutoff: the date its training text stops. It also lists a reliable knowledge cutoff, and context windows in tokens.
- explainerGoogle for Developers. LLMs: What's a large language model?.A language model predicts the next token, like a very large autocomplete. A token can be a word, part of a word or one character. Training needs a huge amount of text. Predictions often contain mistakes.
- documentationAnthropic. Web search tool.Web search lets a model answer with up-to-date information beyond its knowledge cutoff, with citations. It searches for recent news, current prices and scores; for stable facts, maths or writing it answers directly. Search results go into the context window.
Test
You get 5 questions about the whole page. Get 4 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.